With all the problems we have heard about Apple's intelligence lately – late SIRI improvements, summaries of notification of bad news, a generation of uncompromising images, and even more – you might ask yourself what Apple plans to do to put the ship back.
Obviously, new and improved models are important, and therefore in increased training, but Apple has a particularly difficult period, because its privacy policies are much more strict than other companies creating AI products.
In a new article on the Apple automatic learning research site, the company explains a technique that it will use to help its AI be more relevant, more often, without training it on your personal data.
Ensure confidentiality while questioning for use data
Differential confidentiality is a means, as Apple says, “keep an overview of what many Apple users do, while helping to preserve the confidentiality of individual users.”
Basically, each time Apple collects data in a system like this, it first deletes all identification information (device ID, IP address, etc.) slightly alter Data. When millions of users submit results, this “noise” cancels. This is the differential confidentiality part: take enough samples with random noise and deleted identifiers, and you cannot connect specific data to a user.
This is a good way, for example, to obtain a good statistical sample whose emoji are most often chosen, or which automatic correction word is the most used after a particular falsehood – collecting data on the preferences of the user without being really able to trace a particular data point to a user, even if he wanted.
Apple can generate synthetic text representative of common prompts, then use these differential privacy techniques to find out which synthetic samples are most often selected by users. Or to determine which words and sentences are common in Genmoji prompts and what results, users are most likely to choose.
The AI system could generate common sentences used in emails, for example, then send several variants to different users. Then, using differential privacy techniques, Apple can find out which are most frequently selected (while having no capacity to know what an individual has chosen).
Apple has been using this technique for years to collect data intended to improve Quicktype type suggestions, emoji suggestions, search tips, etc. As anonymous as it is, it is still opt. Apple does not collect this type of data, unless you actually activate the analysis of the devices.
Techniques like this are already used to improve Genmoji, and in a next update, they will be used for the generation of images, the image wand, the creation of memories, writing tools and visual intelligence. A Bloomberg report indicates that the new system will be available in a beta update to iOS 18.5, iPados 18.5 and MacOS 18.5 (the second beta version has been published today).
Of course, it is only data collection, and it will take weeks or months of data collection and recycling to improve Apple intelligence features.
